An excluded tool stays excluded.
A request for run_decision_policy was refused by name inside the governed response envelope.
status: ERROR error_code: TOOL_NOT_ALLOWED "run_decision_policy" is not in the EVA MCP allow-listRaw refusal
MCP · verified interoperability
AI can call EVA. EVA still does the analysis.
Kindynos exposes a bounded Model Context Protocol surface so an AI client can discover and invoke approved EVA tools without moving analytical authority into the model. The interface carries requests and results; EVA remains the system that computes the consequence.
Verified 18 Aug 2026Local stdio transport9 approved EVA tools24 acceptance tests passed
ClientAI / agentSelects and invokes an available tool.
→
InteroperabilityMCPGoverned nine-tool allowlist over local stdio.
→
BoundaryEVA tool dispatcherThe same validated service boundary used by EVA Chat.
→
MCP adds no analytical implementation. The live adapter dispatches approved calls and wraps the returned EVA payload in a governance envelope.
Live verification
What was actually demonstrated.
The evidence pack was produced from a live EVA repository and a real MCP server subprocess. Claims below are limited to what was executed or directly inspected.
Tool surface9Approved tools returned by live tools/list.
Acceptance tests24EVA MCP acceptance tests passed in the verification run.
Equivalence17 / 17Top-level fields matched between direct EVA and MCP execution.
Mismatches0The interface changed nothing in the compared analytical result.
Actual EVA call over MCP
A mortgage payoff event, executed rather than narrated.
The reference demonstration used the US housing ecosystem and fired a borrower home-sale payoff event through the live MCP server.
Question represented by the call
What happens to the mortgage chain when the borrower sells the home and pays off the loan?
Tool: run_event_impact
MCP tool arguments
{
"ecosystem_id": "us_housing_v9",
"event_code": "HOU_BORR_HOME_SALE_PAYOFF"
}Loan PV change−$381k
MSR change−$17.6k
MBS pass-through$0 change
Worst stakeholder row−$22.8k
Headline values are rounded for display; the linked response JSON preserves the exact machine output.
The loan and servicing calculations were returned with model/version identity, input/output hashes, run identity and source lineage in the EVA result.
Direct vs MCP
The interface changed nothing analytical.
The same tool was called directly through EVA's validated dispatcher and through the real MCP server. Seventeen top-level fields were compared.
0 mismatches
Verdict: EQUIVALENT. The run ID was byte-identical, each channel-1 asset delta matched, and the input/output hashes matched. Wall-clock timestamps were the only intentionally excluded difference.
Field
Direct EVA
MCP
Match
Scenario
us_housing_v9 × payoffsameYes
Loan delta
−381129.6757923033−381129.6757923033Yes
MSR delta
−17637.17720693946−17637.17720693946Yes
Run ID
engine10::a8d44386…engine10::a8d44386…Yes
Result status
samesameYes
Bounded tool surface
The AI cannot expand the analytical surface by asking for more.
The live EVA MCP surface registers a fixed allowlist of nine tools. Requests outside that surface are returned as structured refusals rather than being improvised by the adapter.
Invalid analytical input remains an EVA refusal.
A bogus event code sent through an allowed tool produced EVA's own catalog error, carried across MCP without being reinterpreted.
status: ERROR error_code: UNKNOWN_EVENT EVT_DOES_NOT_EXIST_XYZRaw refusal
Tool inventory
Nine approved ways into EVA.
The verified surface includes analytical execution, explanation, discovery and glossary access. The adapter does not add a second analytical implementation.
Event analysis
run_event_impact
run_event_impact_with_inputs
Deterministic event consequence execution. The with-inputs variant was registered, schema-verified and unit-tested but not exercised live in this demonstration.
Decision & reaction
run_decision_impact
compute_reaction
Bounded decision consequence and payoff/minimax analysis.
Explain & discover
explain_run
list_ecosystems
list_personas
list_events
define_term
Read stored runs and discover the available analytical vocabulary.
Evidence boundary
What this demonstration does—and does not—establish.
The public claim stays inside the evidence collected on 18 August 2026.
Verified now
EVA MCP ran live over local stdio using the official MCP SDK. Tool discovery, analytical execution, direct-vs-MCP equivalence and typed refusal were exercised.
Prior AI-client evidence
A redacted Claude Desktop log records a real July 2026 client session with tool protocol messages. That client path was evidenced, not re-run during this verification.
Not claimed
No remote or hosted MCP deployment is claimed. EOS MCP was documented in prior build material but was not executed in this verification.
The boundary in one sentence
The AI selects and invokes the tool. EVA performs the authoritative analysis.
MCP is the interoperability layer between them. The analytical result remains an EVA result, with its model identity, run identity, evidence and refusal state intact.
Verified path
AI client
↓
MCP
↓
EVA tool dispatcher
↓
engine10
↓
evidence-bearing result